Remove pass-2 Percolator, flip the pass-2 and pick defaults, reject I/L-isobaric decoys

[brendanx67]

Retires the second-pass Percolator retrain, makes protein-compact and the learned peak pick
the defaults, and rejects I/L-isobaric generated decoys. Mirrors OspreySharp
ProteoWizard/pwiz#4528; the two sides are held
byte-equal by the end-to-end cross-impl comparison, which is the reason for landing them together.

Why pass-2 Percolator goes

It was over optimistic and produced a broken FDP-vs-q curve. Compaction strips most decoys from
the pool the second pass retrains on, so the null is thin and the reported q anti-conservative --
1.57% true FDP at a nominal 1% on Stellar libdecoy entrapment against 0.92% for the first-pass q,
and ~9% on an 82-file SEA-AD set. The linear model trained by the first-pass SVM is now the model
for pass 2 in every mode.

The mode is removed rather than demoted, and an unrecognized OSPREY_PASS2_QVALUE is now an error
raised at startup rather than at the merge node, so a stale sweep script fails in seconds instead
of after Stage 1-5.

Measured on the entrapment oracle, with pass 1 byte-identical across the change so the whole move
is pass 2: on the Stellar libdecoy set, pass-2 true FDP goes from 1.52% to 0.61% at a nominal 1%,
for a 1.6% cost in accepted precursors, and it still adds 9.9% over pass 1 while now measuring
below the line rather than above it. The falling count is not a regression -- a mode that accepts
more at a nominal 1% while measuring 1.5% true FDP is miscalibrated, not better.

protein-compact is not claimed to be fully valid, nor to be the final pass-2 method. It is
target-conditioned and measured anti-conservative at 82 files, and how on-stratum changed peaks
should get their experiment q is still open. Both flipped defaults stay togglable so the arms
remain comparable.

Two correctness fixes that came out of the flip

The changed-peak admission signal was wrong. Admission keyed on presence in the survivor score
override map, and that map is not a "Stage 6 changed" signal -- it holds every post-reconciliation
survivor whose identity resolves, including files Stage 6 never touched, because the effective
path falls back to the original parquet. It is now keyed on a frozen-model score that differs
bit-exactly from the entry's first-pass sidecar score, which is the same discriminator the per-run
q assignment already uses to separate moved from unchanged peaks, and it needs no new plumbing.

Off-stratum peaks re-maxed the experiment accumulator over a subset. They were folded into the
cross-file experiment accumulator but admitted only in the files that changed them, so their
experiment best maxed over a subset of observations. That is not merely asymmetric, it is
guaranteed to understate: reconciliation anchors on the best-scoring peak and corrects the others
toward it, so a changed peak never supplied the maximum. The result inflated those peptides' q and
dropped them. By the same argument the pass-1 experiment q cannot have been invalidated, so it is
now carried rather than recomputed, and only the run-level q is refreshed.

The I/L decoy gate

I and L have identical residue masses, so a decoy differing from a real target only by I<->L is
precursor-mass-identical and produces an identical b/y ladder -- indistinguishable by mass
spectrometry, and detected wherever its target twin is.

The fragment-overlap gate cannot catch this, and that is the part worth keeping: the overlap gate
compares a candidate to its own source target, while a collision is an isobaric match to a
different one. An exact-string audit therefore reports 0 for every library and misses the
population entirely. Simulating the generation path over the 1,390,979-target Astral set with the
overlap gate on: 0 exact collisions, 742 I/L-isobaric ones (0.0534%), e.g. AAEESLR -> LSEEAAR.

The normalised check subsumes the exact one -- a sequence without isoleucine normalises to itself
-- so it replaces rather than supplements it. Carafe already shipped the same check; this closes
the gap on the generation side.

Validation

• cargo fmt --check, cargo clippy -D warnings, cargo test --workspace -- all clean.
• test_collision_check_rejects_a_decoy_isobaric_to_a_different_target asserts the independence
  property -- the overlap gate passes a candidate the collision index rejects -- rather than just
  the rejection, because a test of the rejection alone would pass equally well if the two checks
  were redundant.
• End-to-end cross-impl against OspreySharp, both straight-through, rebuilt binaries:
  Stellar delta=0, and Astral rust=117783 cs=117783 delta=0 with Stage 7 protein FDR and
  blib content both matching at 1e-9. Both implementations moved from 119,088 to 117,783 under the
  two correctness fixes independently, which is what distinguishes one change mirrored correctly
  from two changes that merely resemble each other.

That last point is not decoration: an intermediate state of this branch removed the accumulator
entry without supplying the carry, and the gate immediately reported Rust dropping 619 precursors
against C#. An earlier partial mirror of the admission filter showed up the same way at -152. Both
halves of a pass-2 change have to land together.

Note on scope

OspreySharp additionally puts both flipped defaults into its resume validity key, so a run under
the new defaults cannot adopt cached artifacts written under the old ones. There is no equivalent
change here because this tree has no resume-sidecar or validity-key system.